{"id":9482,"date":"2026-06-21T23:05:20","date_gmt":"2026-06-21T15:05:20","guid":{"rendered":"https:\/\/infernews.com\/blog\/sr-real\/"},"modified":"2026-06-21T23:05:36","modified_gmt":"2026-06-21T15:05:36","slug":"sr-real","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/sr-real\/","title":{"rendered":"SR-REAL \u628a\u7a7a\u9593\u63a8\u7406\u62c6\u6210\u5169\u689d\u8def"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-697ee7ebf5fb.jpg\" alt=\"Repository image for jiyt17\/SR-REAL\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709 spatial VLM \u5f80\u5f80\u7528\u55ae\u4e00\u8def\u7dda\u56de\u7b54\u7a7a\u9593\u554f\u984c\uff0c\u4e0d\u662f\u7d14\u6587\u5b57 chain-of-thought\uff0c\u5c31\u662f\u76f4\u63a5\u9760\u611f\u77e5\u7d50\u679c\u8f38\u51fa\u7b54\u6848\uff1b\u4f5c\u8005\u8a8d\u70ba\u9019\u7a2e\u56fa\u5b9a\u7bc4\u5f0f\u96e3\u4ee5\u540c\u6642\u8655\u7406\u8a9e\u610f\u63a8\u7406\u8207\u7cbe\u78ba\u5e7e\u4f55\u5224\u65b7\u3002SR-REAL \u63d0\u51fa\u7684\u505a\u6cd5\uff0c\u662f\u628a\u7a7a\u9593\u63a8\u7406\u5206\u6210 Language-Only Reasoning\uff08LOR\uff09\u8207 Detect-Then-Reason\uff08DTR\uff09\u5169\u689d\u4e92\u88dc\u8def\u5f91\uff0c\u524d\u8005\u9010\u6b65\u6587\u5b57\u63a8\u7406\uff0c\u5f8c\u8005\u5148\u627e 3D \u5e7e\u4f55\u7dda\u7d22\uff0c\u518d\u505a\u660e\u78ba\u5e7e\u4f55\u63a8\u65b7\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u5c6c\u65bc<strong>\u6846\u67b6\u52a0\u8a13\u7df4\u6d41\u7a0b\u5be6\u4f5c<\/strong>\uff0c\u6838\u5fc3\u662f\u5f37\u5316 spatial vision-language models \u5728\u8907\u96dc\u7a7a\u9593\u554f\u7b54\u4e2d\u7684\u5224\u65b7\u80fd\u529b\u3002\u5b83\u4e0d\u662f\u55ae\u7d14\u65b0\u589e\u8cc7\u6599\u96c6\uff0c\u800c\u662f\u5f9e cold-start supervised fine-tuning \u5230 reinforcement learning\uff08RL\uff09\u90fd\u91cd\u65b0\u5b89\u6392\uff0c\u4e26\u52a0\u5165 region-to-3D \u4ecb\u9762\uff0c\u4ee4\u6a21\u578b\u53ef\u628a region tokens \u9023\u5230 3D \u5ea7\u6a19\u3001\u4e2d\u5fc3\u9ede\u6216 bounding boxes\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SR-REAL \u91cd\u9ede\u96c6\u4e2d\u5728\u8cc7\u6599\u6e96\u5099\u8207\u8a13\u7df4\u524d\u8655\u7406\u3002\u6d41\u7a0b\u4e0a\u6703\u5148\u7528 SPAR\u3001EmbodiedScan \u7b49\u4f86\u6e90\u6574\u7406\u7269\u4ef6\u5c0d\u61c9\u8207 3D \u5ea7\u6a19\uff0c\u518d\u7531 expert.py \u751f\u6210\u63a8\u7406\u93c8\uff0c\u914d\u5408 qwen3.py \u62bd\u53d6\u7269\u4ef6\u540d\u7a31\uff0c\u6700\u5f8c\u7d44\u6210 DTR \u6307\u4ee4\u5fae\u8abf\u8cc7\u6599\uff1b\u82e5\u4e0d\u60f3\u81ea\u884c\u91cd\u5efa\uff0c\u4e5f\u53ef\u76f4\u63a5\u4e0b\u8f09\u4f5c\u8005\u5df2\u6574\u7406\u597d\u7684 Hugging Face \u6578\u64da\u3002\u9019\u8868\u793a\u5b83\u8f03\u9069\u5408\u6709 Python\u3001\u8cc7\u6599\u8655\u7406\u53ca\u591a\u6a21\u614b\u8a13\u7df4\u57fa\u790e\u7684\u7814\u7a76\u5718\u968a\uff0c\u800c\u4e0d\u662f\u5373\u88dd\u5373\u7528\u7684\u7d42\u7aef\u5de5\u5177\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u548c\u540c\u985e\u505a\u6cd5\u76f8\u6bd4\uff0cSR-REAL \u4e0d\u5047\u8a2d\u6240\u6709\u7a7a\u9593\u554f\u984c\u90fd\u61c9\u8a72\u7528\u540c\u4e00\u7a2e reasoning path\u3002\u4f5c\u8005\u7684\u53d6\u5411\u5f88\u6e05\u695a\uff1a\u8a9e\u610f\u95dc\u4fc2\u9069\u5408 LOR\uff0c\u6d89\u53ca\u660e\u78ba\u4f4d\u7f6e\u3001\u8ddd\u96e2\u3001\u4e2d\u5fc3\u9ede\u3001\u6846\u9078\u5340\u57df\u7684\u984c\u76ee\u5247\u4ea4\u7d66 DTR\uff1b\u4ee3\u50f9\u662f\u6574\u500b\u8cc7\u6599\u69cb\u5efa\u8207\u8a13\u7df4\u6d41\u7a0b\u66f4\u8907\u96dc\uff0c\u5c0d grounding \u8cc7\u6599\u54c1\u8cea\u4ea6\u66f4\u654f\u611f\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u91cd\u9ede\u4e0d\u5728\u55ae\u4e00\u6a21\u578b\u7d50\u69cb\uff0c\u800c\u5728 <strong>LOR + DTR<\/strong> \u96d9\u8def\u5f91\u63a8\u7406\u8a2d\u8a08<\/li>\n\n\n\n<li>DTR \u6703\u5148\u8655\u7406 region tokens \u8207 3D \u5e7e\u4f55\u7dda\u7d22\uff0c\u518d\u505a\u7a7a\u9593\u5224\u65b7<\/li>\n\n\n\n<li>\u8a13\u7df4\u5206\u70ba cold-start supervised fine-tuning \u8207 reinforcement learning\uff08RL\uff09\u5169\u6bb5<\/li>\n\n\n\n<li>\u5df2\u63d0\u53ca accuracy\u3001format\u3001detection rewards\uff0c\u986f\u793a\u8a55\u6e2c\u4e0d\u53ea\u770b\u7b54\u5c0d\u8207\u5426\uff0c\u4e5f\u770b\u8f38\u51fa\u683c\u5f0f\u53ca\u5e7e\u4f55\u5c0d\u9f4a<\/li>\n\n\n\n<li>\u76f8\u95dc\u6a21\u578b\u8207\u8cc7\u6599\u4f86\u6e90\u5305\u62ec spatial VLM\u3001SR-3D\u3001Qwen3\u3001SPAR\u3001EmbodiedScan\u3001SpatialRGPT\u3001Omni3D\u3001CA1M\u3001OmniNOCS<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">SR-REAL \u5728\u591a\u500b spatial benchmarks \u6709\u660e\u986f\u63d0\u5347\uff0c\u4e26\u5f37\u8abf\u55ae\u4e00 RL-trained model \u53ef\u540c\u6642\u652f\u63f4\u5169\u689d\u8def\u5f91\uff0c\u4e14\u4e0d\u7528 per-task tuning \u4e5f\u80fd\u8de8\u8cc7\u6599\u96c6\u6cdb\u5316\u3002\u4e0d\u904e\u5132\u5b58\u5eab\u7247\u6bb5\u672a\u5b8c\u6574\u5217\u51fa\u8a73\u7d30\u5206\u6578\u8207\u5c0d\u7167\u8868\uff0c\u56e0\u6b64\u8f03\u7a69\u59a5\u7684\u5224\u65b7\u662f\uff1a\u9019\u662f\u4e00\u500b\u7814\u7a76\u5473\u5f88\u91cd\u3001\u65b9\u6cd5\u8ad6\u6e05\u6670\u7684\u9805\u76ee\uff0c\u9069\u5408\u95dc\u6ce8 spatial reasoning\u30013D grounding\u3001multimodal instruction tuning \u7684\u5718\u968a\u62ff\u4f86\u91cd\u73fe\u8207\u5ef6\u4f38\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/jiyt17\/SR-REAL\" rel=\"noopener noreferrer\">https:\/\/github.com\/jiyt17\/SR-REAL<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9805\u76ee\u4e3b\u9801\uff1a<\/strong> <a href=\"https:\/\/sr-real.github.io\/\" rel=\"noopener noreferrer\">https:\/\/sr-real.github.io\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>SR-REAL \u4e0d\u662f\u53ea\u9760\u6587\u5b57\u63a8\u7406\uff0c\u4e5f\u52a0\u5165 3D \u5b9a\u4f4d\u518d\u5224\u65b7\u3002\u5b83\u60f3\u89e3\u6c7a\u7a7a\u9593 VLM \u9047\u5230\u8ddd\u96e2\u3001\u65b9\u4f4d\u3001\u591a\u6b65\u63a8\u7406\u6642\u5e38\u51fa\u932f\u7684\u554f\u984c\u3002<\/p>\n","protected":false},"author":8,"featured_media":9481,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[185,163,169,178,179,147,153,116,38,132,152,105,119,150,76,127,146,197],"tags":[],"class_list":["post-9482","post","type-post","status-publish","format-standard","hentry","category-qwen","category-163","category-169","category-google","category-nvidia","category-deepseek","category-openai","category-agentic","category-38","category-3d","category-python","category-python-nlp","category-119","category-150","category-76","category-127","category-146","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9482","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=9482"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9482\/revisions"}],"predecessor-version":[{"id":9484,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9482\/revisions\/9484"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9481"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9482"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9482"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9482"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}